bcachefs: Reorganize extents.c

Signed-off-by: Kent Overstreet <kent.overstreet@gmail.com>
Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>

fs/bcachefs/extent_update.c

@@ -200,7 +200,7 @@ bch2_extent_can_insert(struct btree_trans *trans,
 		*u64s += _k->u64s;
 
 	if (overlap == BCH_EXTENT_OVERLAP_MIDDLE &&
-	    (sectors = bch2_extent_is_compressed(k))) {
+	    (sectors = bch2_bkey_sectors_compressed(k))) {
 		int flags = trans->flags & BTREE_INSERT_NOFAIL
 			? BCH_DISK_RESERVATION_NOFAIL : 0;
 

fs/bcachefs/extents.c

@@ -24,81 +24,15 @@
 #include "trace.h"
 #include "util.h"
 
-unsigned bch2_bkey_nr_ptrs(struct bkey_s_c k)
-{
-	struct bkey_ptrs_c p = bch2_bkey_ptrs_c(k);
-	const struct bch_extent_ptr *ptr;
-	unsigned nr_ptrs = 0;
-
-	bkey_for_each_ptr(p, ptr)
-		nr_ptrs++;
-
-	return nr_ptrs;
-}
-
-unsigned bch2_bkey_nr_dirty_ptrs(struct bkey_s_c k)
-{
-	unsigned nr_ptrs = 0;
-
-	switch (k.k->type) {
-	case KEY_TYPE_btree_ptr:
-	case KEY_TYPE_extent:
-	case KEY_TYPE_reflink_v: {
-		struct bkey_ptrs_c p = bch2_bkey_ptrs_c(k);
-		const struct bch_extent_ptr *ptr;
-
-		bkey_for_each_ptr(p, ptr)
-			nr_ptrs += !ptr->cached;
-		BUG_ON(!nr_ptrs);
-		break;
-	}
-	case KEY_TYPE_reservation:
-		nr_ptrs = bkey_s_c_to_reservation(k).v->nr_replicas;
-		break;
-	}
-
-	return nr_ptrs;
-}
-
-static unsigned bch2_extent_ptr_durability(struct bch_fs *c,
-					   struct extent_ptr_decoded p)
-{
-	unsigned durability = 0;
-	struct bch_dev *ca;
-
-	if (p.ptr.cached)
-		return 0;
-
-	ca = bch_dev_bkey_exists(c, p.ptr.dev);
-
-	if (ca->mi.state != BCH_MEMBER_STATE_FAILED)
-		durability = max_t(unsigned, durability, ca->mi.durability);
-
-	if (p.has_ec) {
-		struct stripe *s =
-			genradix_ptr(&c->stripes[0], p.ec.idx);
-
-		if (WARN_ON(!s))
-			goto out;
-
-		durability = max_t(unsigned, durability, s->nr_redundant);
-	}
-out:
-	return durability;
-}
-
-unsigned bch2_bkey_durability(struct bch_fs *c, struct bkey_s_c k)
-{
-	struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
-	const union bch_extent_entry *entry;
-	struct extent_ptr_decoded p;
-	unsigned durability = 0;
-
-	bkey_for_each_ptr_decode(k.k, ptrs, p, entry)
-		durability += bch2_extent_ptr_durability(c, p);
+static unsigned bch2_crc_field_size_max[] = {
+	[BCH_EXTENT_ENTRY_crc32] = CRC32_SIZE_MAX,
+	[BCH_EXTENT_ENTRY_crc64] = CRC64_SIZE_MAX,
+	[BCH_EXTENT_ENTRY_crc128] = CRC128_SIZE_MAX,
+};
 
-	return durability;
-}
+static void bch2_extent_crc_pack(union bch_extent_crc *,
+				 struct bch_extent_crc_unpacked,
+				 enum bch_extent_entry_type);
 
 static struct bch_dev_io_failures *dev_io_failures(struct bch_io_failures *f,
 						   unsigned dev)
@@ -218,172 +152,299 @@ int bch2_bkey_pick_read_device(struct bch_fs *c, struct bkey_s_c k,
 	return ret;
 }
 
-void bch2_bkey_append_ptr(struct bkey_i *k,
-			  struct bch_extent_ptr ptr)
-{
-	EBUG_ON(bch2_bkey_has_device(bkey_i_to_s_c(k), ptr.dev));
-
-	switch (k->k.type) {
-	case KEY_TYPE_btree_ptr:
-	case KEY_TYPE_extent:
-		EBUG_ON(bkey_val_u64s(&k->k) >= BKEY_EXTENT_VAL_U64s_MAX);
-
-		ptr.type = 1 << BCH_EXTENT_ENTRY_ptr;
-
-		memcpy((void *) &k->v + bkey_val_bytes(&k->k),
-		       &ptr,
-		       sizeof(ptr));
-		k->u64s++;
-		break;
-	default:
-		BUG();
-	}
-}
+/* KEY_TYPE_btree_ptr: */
 
-void bch2_bkey_drop_device(struct bkey_s k, unsigned dev)
+const char *bch2_btree_ptr_invalid(const struct bch_fs *c, struct bkey_s_c k)
 {
-	struct bch_extent_ptr *ptr;
+	if (bkey_val_u64s(k.k) > BKEY_BTREE_PTR_VAL_U64s_MAX)
+		return "value too big";
 
-	bch2_bkey_drop_ptrs(k, ptr, ptr->dev == dev);
+	return bch2_bkey_ptrs_invalid(c, k);
 }
 
-const struct bch_extent_ptr *
-bch2_bkey_has_device(struct bkey_s_c k, unsigned dev)
+void bch2_btree_ptr_debugcheck(struct bch_fs *c, struct bkey_s_c k)
 {
 	struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
 	const struct bch_extent_ptr *ptr;
+	const char *err;
+	char buf[160];
+	struct bucket_mark mark;
+	struct bch_dev *ca;
 
-	bkey_for_each_ptr(ptrs, ptr)
-		if (ptr->dev == dev)
-			return ptr;
-
-	return NULL;
-}
-
-bool bch2_bkey_has_target(struct bch_fs *c, struct bkey_s_c k, unsigned target)
-{
-	struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
-	const struct bch_extent_ptr *ptr;
+	bch2_fs_bug_on(!test_bit(BCH_FS_REBUILD_REPLICAS, &c->flags) &&
+		       !bch2_bkey_replicas_marked(c, k, false), c,
+		       "btree key bad (replicas not marked in superblock):\n%s",
+		       (bch2_bkey_val_to_text(&PBUF(buf), c, k), buf));
 
-	bkey_for_each_ptr(ptrs, ptr)
-		if (bch2_dev_in_target(c, ptr->dev, target) &&
-		    (!ptr->cached ||
-		     !ptr_stale(bch_dev_bkey_exists(c, ptr->dev), ptr)))
-			return true;
+	if (!test_bit(BCH_FS_INITIAL_GC_DONE, &c->flags))
+		return;
 
-	return false;
-}
+	bkey_for_each_ptr(ptrs, ptr) {
+		ca = bch_dev_bkey_exists(c, ptr->dev);
 
-/* extent specific utility code */
+		mark = ptr_bucket_mark(ca, ptr);
 
-const struct bch_extent_ptr *
-bch2_extent_has_device(struct bkey_s_c_extent e, unsigned dev)
-{
-	const struct bch_extent_ptr *ptr;
+		err = "stale";
+		if (gen_after(mark.gen, ptr->gen))
+			goto err;
 
-	extent_for_each_ptr(e, ptr)
-		if (ptr->dev == dev)
-			return ptr;
+		err = "inconsistent";
+		if (mark.data_type != BCH_DATA_BTREE ||
+		    mark.dirty_sectors < c->opts.btree_node_size)
+			goto err;
+	}
 
-	return NULL;
+	return;
+err:
+	bch2_bkey_val_to_text(&PBUF(buf), c, k);
+	bch2_fs_bug(c, "%s btree pointer %s: bucket %zi gen %i mark %08x",
+		    err, buf, PTR_BUCKET_NR(ca, ptr),
+		    mark.gen, (unsigned) mark.v.counter);
 }
 
-const struct bch_extent_ptr *
-bch2_extent_has_group(struct bch_fs *c, struct bkey_s_c_extent e, unsigned group)
+void bch2_btree_ptr_to_text(struct printbuf *out, struct bch_fs *c,
+			    struct bkey_s_c k)
 {
-	const struct bch_extent_ptr *ptr;
-
-	extent_for_each_ptr(e, ptr) {
-		struct bch_dev *ca = bch_dev_bkey_exists(c, ptr->dev);
+	bch2_bkey_ptrs_to_text(out, c, k);
+}
 
-		if (ca->mi.group &&
-		    ca->mi.group - 1 == group)
-			return ptr;
-	}
+/* KEY_TYPE_extent: */
 
-	return NULL;
+const char *bch2_extent_invalid(const struct bch_fs *c, struct bkey_s_c k)
+{
+	return bch2_bkey_ptrs_invalid(c, k);
 }
 
-unsigned bch2_extent_is_compressed(struct bkey_s_c k)
+void bch2_extent_debugcheck(struct bch_fs *c, struct bkey_s_c k)
 {
-	struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
+	struct bkey_s_c_extent e = bkey_s_c_to_extent(k);
 	const union bch_extent_entry *entry;
 	struct extent_ptr_decoded p;
-	unsigned ret = 0;
+	char buf[160];
 
-	bkey_for_each_ptr_decode(k.k, ptrs, p, entry)
-		if (!p.ptr.cached &&
-		    p.crc.compression_type != BCH_COMPRESSION_NONE)
-			ret += p.crc.compressed_size;
+	/*
+	 * XXX: we should be doing most/all of these checks at startup time,
+	 * where we check bch2_bkey_invalid() in btree_node_read_done()
+	 *
+	 * But note that we can't check for stale pointers or incorrect gc marks
+	 * until after journal replay is done (it might be an extent that's
+	 * going to get overwritten during replay)
+	 */
 
-	return ret;
-}
+	if (percpu_down_read_trylock(&c->mark_lock)) {
+		bch2_fs_bug_on(!test_bit(BCH_FS_REBUILD_REPLICAS, &c->flags) &&
+			       !bch2_bkey_replicas_marked_locked(c, e.s_c, false), c,
+			       "extent key bad (replicas not marked in superblock):\n%s",
+			       (bch2_bkey_val_to_text(&PBUF(buf), c, e.s_c), buf));
+		percpu_up_read(&c->mark_lock);
+	}
+	/*
+	 * If journal replay hasn't finished, we might be seeing keys
+	 * that will be overwritten by the time journal replay is done:
+	 */
+	if (!test_bit(JOURNAL_REPLAY_DONE, &c->journal.flags))
+		return;
 
-bool bch2_bkey_matches_ptr(struct bch_fs *c, struct bkey_s_c k,
-			   struct bch_extent_ptr m, u64 offset)
-{
-	struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
-	const union bch_extent_entry *entry;
-	struct extent_ptr_decoded p;
+	extent_for_each_ptr_decode(e, p, entry) {
+		struct bch_dev *ca	= bch_dev_bkey_exists(c, p.ptr.dev);
+		struct bucket_mark mark = ptr_bucket_mark(ca, &p.ptr);
+		unsigned stale		= gen_after(mark.gen, p.ptr.gen);
+		unsigned disk_sectors	= ptr_disk_sectors(p);
+		unsigned mark_sectors	= p.ptr.cached
+			? mark.cached_sectors
+			: mark.dirty_sectors;
 
-	bkey_for_each_ptr_decode(k.k, ptrs, p, entry)
-		if (p.ptr.dev	== m.dev &&
-		    p.ptr.gen	== m.gen &&
-		    (s64) p.ptr.offset + p.crc.offset - bkey_start_offset(k.k) ==
-		    (s64) m.offset  - offset)
-			return true;
+		bch2_fs_bug_on(stale && !p.ptr.cached, c,
+			       "stale dirty pointer (ptr gen %u bucket %u",
+			       p.ptr.gen, mark.gen);
 
-	return false;
+		bch2_fs_bug_on(stale > 96, c, "key too stale: %i", stale);
+
+		bch2_fs_bug_on(!stale &&
+			       (mark.data_type != BCH_DATA_USER ||
+				mark_sectors < disk_sectors), c,
+			       "extent pointer not marked: %s:\n"
+			       "type %u sectors %u < %u",
+			       (bch2_bkey_val_to_text(&PBUF(buf), c, e.s_c), buf),
+			       mark.data_type,
+			       mark_sectors, disk_sectors);
+	}
 }
 
-static union bch_extent_entry *extent_entry_prev(struct bkey_ptrs ptrs,
-					  union bch_extent_entry *entry)
+void bch2_extent_to_text(struct printbuf *out, struct bch_fs *c,
+			 struct bkey_s_c k)
 {
-	union bch_extent_entry *i = ptrs.start;
-
-	if (i == entry)
-		return NULL;
-
-	while (extent_entry_next(i) != entry)
-		i = extent_entry_next(i);
-	return i;
+	bch2_bkey_ptrs_to_text(out, c, k);
 }
 
-union bch_extent_entry *bch2_bkey_drop_ptr(struct bkey_s k,
-					   struct bch_extent_ptr *ptr)
+enum merge_result bch2_extent_merge(struct bch_fs *c,
+				    struct bkey_s _l, struct bkey_s _r)
 {
-	struct bkey_ptrs ptrs = bch2_bkey_ptrs(k);
-	union bch_extent_entry *dst, *src, *prev;
-	bool drop_crc = true;
-
-	EBUG_ON(ptr < &ptrs.start->ptr ||
-		ptr >= &ptrs.end->ptr);
-	EBUG_ON(ptr->type != 1 << BCH_EXTENT_ENTRY_ptr);
+	struct bkey_s_extent l = bkey_s_to_extent(_l);
+	struct bkey_s_extent r = bkey_s_to_extent(_r);
+	union bch_extent_entry *en_l = l.v->start;
+	union bch_extent_entry *en_r = r.v->start;
+	struct bch_extent_crc_unpacked crc_l, crc_r;
 
-	src = extent_entry_next(to_entry(ptr));
-	if (src != ptrs.end &&
-	    !extent_entry_is_crc(src))
-		drop_crc = false;
+	if (bkey_val_u64s(l.k) != bkey_val_u64s(r.k))
+		return BCH_MERGE_NOMERGE;
 
-	dst = to_entry(ptr);
-	while ((prev = extent_entry_prev(ptrs, dst))) {
-		if (extent_entry_is_ptr(prev))
+	crc_l = bch2_extent_crc_unpack(l.k, NULL);
+
+	extent_for_each_entry(l, en_l) {
+		en_r = vstruct_idx(r.v, (u64 *) en_l - l.v->_data);
+
+		if (extent_entry_type(en_l) != extent_entry_type(en_r))
+			return BCH_MERGE_NOMERGE;
+
+		switch (extent_entry_type(en_l)) {
+		case BCH_EXTENT_ENTRY_ptr: {
+			const struct bch_extent_ptr *lp = &en_l->ptr;
+			const struct bch_extent_ptr *rp = &en_r->ptr;
+			struct bch_dev *ca;
+
+			if (lp->offset + crc_l.compressed_size != rp->offset ||
+			    lp->dev			!= rp->dev ||
+			    lp->gen			!= rp->gen)
+				return BCH_MERGE_NOMERGE;
+
+			/* We don't allow extents to straddle buckets: */
+			ca = bch_dev_bkey_exists(c, lp->dev);
+
+			if (PTR_BUCKET_NR(ca, lp) != PTR_BUCKET_NR(ca, rp))
+				return BCH_MERGE_NOMERGE;
+
+			break;
+		}
+		case BCH_EXTENT_ENTRY_stripe_ptr:
+			if (en_l->stripe_ptr.block	!= en_r->stripe_ptr.block ||
+			    en_l->stripe_ptr.idx	!= en_r->stripe_ptr.idx)
+				return BCH_MERGE_NOMERGE;
 			break;
+		case BCH_EXTENT_ENTRY_crc32:
+		case BCH_EXTENT_ENTRY_crc64:
+		case BCH_EXTENT_ENTRY_crc128:
+			crc_l = bch2_extent_crc_unpack(l.k, entry_to_crc(en_l));
+			crc_r = bch2_extent_crc_unpack(r.k, entry_to_crc(en_r));
+
+			if (crc_l.csum_type		!= crc_r.csum_type ||
+			    crc_l.compression_type	!= crc_r.compression_type ||
+			    crc_l.nonce			!= crc_r.nonce)
+				return BCH_MERGE_NOMERGE;
+
+			if (crc_l.offset + crc_l.live_size != crc_l.compressed_size ||
+			    crc_r.offset)
+				return BCH_MERGE_NOMERGE;
+
+			if (!bch2_checksum_mergeable(crc_l.csum_type))
+				return BCH_MERGE_NOMERGE;
+
+			if (crc_l.compression_type)
+				return BCH_MERGE_NOMERGE;
+
+			if (crc_l.csum_type &&
+			    crc_l.uncompressed_size +
+			    crc_r.uncompressed_size > c->sb.encoded_extent_max)
+				return BCH_MERGE_NOMERGE;
+
+			if (crc_l.uncompressed_size + crc_r.uncompressed_size - 1 >
+			    bch2_crc_field_size_max[extent_entry_type(en_l)])
+				return BCH_MERGE_NOMERGE;
 
-		if (extent_entry_is_crc(prev)) {
-			if (drop_crc)
-				dst = prev;
 			break;
+		default:
+			return BCH_MERGE_NOMERGE;
+		}
 	}
 
-		dst = prev;
+	extent_for_each_entry(l, en_l) {
+		struct bch_extent_crc_unpacked crc_l, crc_r;
+
+		en_r = vstruct_idx(r.v, (u64 *) en_l - l.v->_data);
+
+		if (!extent_entry_is_crc(en_l))
+			continue;
+
+		crc_l = bch2_extent_crc_unpack(l.k, entry_to_crc(en_l));
+		crc_r = bch2_extent_crc_unpack(r.k, entry_to_crc(en_r));
+
+		crc_l.csum = bch2_checksum_merge(crc_l.csum_type,
+						 crc_l.csum,
+						 crc_r.csum,
+						 crc_r.uncompressed_size << 9);
+
+		crc_l.uncompressed_size	+= crc_r.uncompressed_size;
+		crc_l.compressed_size	+= crc_r.compressed_size;
+
+		bch2_extent_crc_pack(entry_to_crc(en_l), crc_l,
+				     extent_entry_type(en_l));
 	}
 
-	memmove_u64s_down(dst, src,
-			  (u64 *) ptrs.end - (u64 *) src);
-	k.k->u64s -= (u64 *) src - (u64 *) dst;
+	bch2_key_resize(l.k, l.k->size + r.k->size);
 
-	return dst;
+	return BCH_MERGE_MERGE;
+}
+
+/* KEY_TYPE_reservation: */
+
+const char *bch2_reservation_invalid(const struct bch_fs *c, struct bkey_s_c k)
+{
+	struct bkey_s_c_reservation r = bkey_s_c_to_reservation(k);
+
+	if (bkey_val_bytes(k.k) != sizeof(struct bch_reservation))
+		return "incorrect value size";
+
+	if (!r.v->nr_replicas || r.v->nr_replicas > BCH_REPLICAS_MAX)
+		return "invalid nr_replicas";
+
+	return NULL;
+}
+
+void bch2_reservation_to_text(struct printbuf *out, struct bch_fs *c,
+			      struct bkey_s_c k)
+{
+	struct bkey_s_c_reservation r = bkey_s_c_to_reservation(k);
+
+	pr_buf(out, "generation %u replicas %u",
+	       le32_to_cpu(r.v->generation),
+	       r.v->nr_replicas);
+}
+
+enum merge_result bch2_reservation_merge(struct bch_fs *c,
+					 struct bkey_s _l, struct bkey_s _r)
+{
+	struct bkey_s_reservation l = bkey_s_to_reservation(_l);
+	struct bkey_s_reservation r = bkey_s_to_reservation(_r);
+
+	if (l.v->generation != r.v->generation ||
+	    l.v->nr_replicas != r.v->nr_replicas)
+		return BCH_MERGE_NOMERGE;
+
+	if ((u64) l.k->size + r.k->size > KEY_SIZE_MAX) {
+		bch2_key_resize(l.k, KEY_SIZE_MAX);
+		bch2_cut_front_s(l.k->p, r.s);
+		return BCH_MERGE_PARTIAL;
+	}
+
+	bch2_key_resize(l.k, l.k->size + r.k->size);
+
+	return BCH_MERGE_MERGE;
+}
+
+/* Extent checksum entries: */
+
+/* returns true if not equal */
+static inline bool bch2_crc_unpacked_cmp(struct bch_extent_crc_unpacked l,
+					 struct bch_extent_crc_unpacked r)
+{
+	return (l.csum_type		!= r.csum_type ||
+		l.compression_type	!= r.compression_type ||
+		l.compressed_size	!= r.compressed_size ||
+		l.uncompressed_size	!= r.uncompressed_size ||
+		l.offset		!= r.offset ||
+		l.live_size		!= r.live_size ||
+		l.nonce			!= r.nonce ||
+		bch2_crc_cmp(l.csum, r.csum));
 }
 
 static inline bool can_narrow_crc(struct bch_extent_crc_unpacked u,
@@ -462,509 +523,237 @@ bool bch2_bkey_narrow_crcs(struct bkey_i *k, struct bch_extent_crc_unpacked n)
 	return ret;
 }
 
-/* returns true if not equal */
-static inline bool bch2_crc_unpacked_cmp(struct bch_extent_crc_unpacked l,
-					 struct bch_extent_crc_unpacked r)
-{
-	return (l.csum_type		!= r.csum_type ||
-		l.compression_type	!= r.compression_type ||
-		l.compressed_size	!= r.compressed_size ||
-		l.uncompressed_size	!= r.uncompressed_size ||
-		l.offset		!= r.offset ||
-		l.live_size		!= r.live_size ||
-		l.nonce			!= r.nonce ||
-		bch2_crc_cmp(l.csum, r.csum));
-}
-
-void bch2_ptr_swab(const struct bkey_format *f, struct bkey_packed *k)
+static void bch2_extent_crc_pack(union bch_extent_crc *dst,
+				 struct bch_extent_crc_unpacked src,
+				 enum bch_extent_entry_type type)
 {
-	union bch_extent_entry *entry;
-	u64 *d = (u64 *) bkeyp_val(f, k);
-	unsigned i;
-
-	for (i = 0; i < bkeyp_val_u64s(f, k); i++)
-		d[i] = swab64(d[i]);
+#define set_common_fields(_dst, _src)					\
+		_dst.type		= 1 << type;			\
+		_dst.csum_type		= _src.csum_type,		\
+		_dst.compression_type	= _src.compression_type,	\
+		_dst._compressed_size	= _src.compressed_size - 1,	\
+		_dst._uncompressed_size	= _src.uncompressed_size - 1,	\
+		_dst.offset		= _src.offset
 
-	for (entry = (union bch_extent_entry *) d;
-	     entry < (union bch_extent_entry *) (d + bkeyp_val_u64s(f, k));
-	     entry = extent_entry_next(entry)) {
-		switch (extent_entry_type(entry)) {
-		case BCH_EXTENT_ENTRY_ptr:
-			break;
+	switch (type) {
 	case BCH_EXTENT_ENTRY_crc32:
-			entry->crc32.csum = swab32(entry->crc32.csum);
+		set_common_fields(dst->crc32, src);
+		dst->crc32.csum	 = *((__le32 *) &src.csum.lo);
 		break;
 	case BCH_EXTENT_ENTRY_crc64:
-			entry->crc64.csum_hi = swab16(entry->crc64.csum_hi);
-			entry->crc64.csum_lo = swab64(entry->crc64.csum_lo);
+		set_common_fields(dst->crc64, src);
+		dst->crc64.nonce	= src.nonce;
+		dst->crc64.csum_lo	= src.csum.lo;
+		dst->crc64.csum_hi	= *((__le16 *) &src.csum.hi);
 		break;
 	case BCH_EXTENT_ENTRY_crc128:
-			entry->crc128.csum.hi = (__force __le64)
-				swab64((__force u64) entry->crc128.csum.hi);
-			entry->crc128.csum.lo = (__force __le64)
-				swab64((__force u64) entry->crc128.csum.lo);
-			break;
-		case BCH_EXTENT_ENTRY_stripe_ptr:
+		set_common_fields(dst->crc128, src);
+		dst->crc128.nonce	= src.nonce;
+		dst->crc128.csum	= src.csum;
 		break;
+	default:
+		BUG();
 	}
-	}
+#undef set_common_fields
 }
 
-void bch2_bkey_ptrs_to_text(struct printbuf *out, struct bch_fs *c,
-			    struct bkey_s_c k)
+void bch2_extent_crc_append(struct bkey_i *k,
+			    struct bch_extent_crc_unpacked new)
 {
-	struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
-	const union bch_extent_entry *entry;
-	struct bch_extent_crc_unpacked crc;
-	const struct bch_extent_ptr *ptr;
-	const struct bch_extent_stripe_ptr *ec;
-	struct bch_dev *ca;
-	bool first = true;
+	struct bkey_ptrs ptrs = bch2_bkey_ptrs(bkey_i_to_s(k));
+	union bch_extent_crc *crc = (void *) ptrs.end;
+	enum bch_extent_entry_type type;
 
-	bkey_extent_entry_for_each(ptrs, entry) {
-		if (!first)
-			pr_buf(out, " ");
+	if (bch_crc_bytes[new.csum_type]	<= 4 &&
+	    new.uncompressed_size - 1		<= CRC32_SIZE_MAX &&
+	    new.nonce				<= CRC32_NONCE_MAX)
+		type = BCH_EXTENT_ENTRY_crc32;
+	else if (bch_crc_bytes[new.csum_type]	<= 10 &&
+		   new.uncompressed_size - 1	<= CRC64_SIZE_MAX &&
+		   new.nonce			<= CRC64_NONCE_MAX)
+		type = BCH_EXTENT_ENTRY_crc64;
+	else if (bch_crc_bytes[new.csum_type]	<= 16 &&
+		   new.uncompressed_size - 1	<= CRC128_SIZE_MAX &&
+		   new.nonce			<= CRC128_NONCE_MAX)
+		type = BCH_EXTENT_ENTRY_crc128;
+	else
+		BUG();
 
-		switch (__extent_entry_type(entry)) {
-		case BCH_EXTENT_ENTRY_ptr:
-			ptr = entry_to_ptr(entry);
-			ca = ptr->dev < c->sb.nr_devices && c->devs[ptr->dev]
-				? bch_dev_bkey_exists(c, ptr->dev)
-				: NULL;
+	bch2_extent_crc_pack(crc, new, type);
 
-			pr_buf(out, "ptr: %u:%llu gen %u%s%s", ptr->dev,
-			       (u64) ptr->offset, ptr->gen,
-			       ptr->cached ? " cached" : "",
-			       ca && ptr_stale(ca, ptr)
-			       ? " stale" : "");
-			break;
-		case BCH_EXTENT_ENTRY_crc32:
-		case BCH_EXTENT_ENTRY_crc64:
-		case BCH_EXTENT_ENTRY_crc128:
-			crc = bch2_extent_crc_unpack(k.k, entry_to_crc(entry));
+	k->k.u64s += extent_entry_u64s(ptrs.end);
 
-			pr_buf(out, "crc: c_size %u size %u offset %u nonce %u csum %u compress %u",
-			       crc.compressed_size,
-			       crc.uncompressed_size,
-			       crc.offset, crc.nonce,
-			       crc.csum_type,
-			       crc.compression_type);
-			break;
-		case BCH_EXTENT_ENTRY_stripe_ptr:
-			ec = &entry->stripe_ptr;
+	EBUG_ON(bkey_val_u64s(&k->k) > BKEY_EXTENT_VAL_U64s_MAX);
+}
 
-			pr_buf(out, "ec: idx %llu block %u",
-			       (u64) ec->idx, ec->block);
-			break;
-		default:
-			pr_buf(out, "(invalid extent entry %.16llx)", *((u64 *) entry));
-			return;
-		}
+/* Generic code for keys with pointers: */
 
-		first = false;
-	}
+unsigned bch2_bkey_nr_ptrs(struct bkey_s_c k)
+{
+	return bch2_bkey_devs(k).nr;
 }
 
-static const char *extent_ptr_invalid(const struct bch_fs *c,
-				      struct bkey_s_c k,
-				      const struct bch_extent_ptr *ptr,
-				      unsigned size_ondisk,
-				      bool metadata)
+unsigned bch2_bkey_nr_ptrs_allocated(struct bkey_s_c k)
 {
-	struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
-	const struct bch_extent_ptr *ptr2;
-	struct bch_dev *ca;
-
-	if (!bch2_dev_exists2(c, ptr->dev))
-		return "pointer to invalid device";
-
-	ca = bch_dev_bkey_exists(c, ptr->dev);
-	if (!ca)
-		return "pointer to invalid device";
-
-	bkey_for_each_ptr(ptrs, ptr2)
-		if (ptr != ptr2 && ptr->dev == ptr2->dev)
-			return "multiple pointers to same device";
-
-	if (ptr->offset + size_ondisk > bucket_to_sector(ca, ca->mi.nbuckets))
-		return "offset past end of device";
-
-	if (ptr->offset < bucket_to_sector(ca, ca->mi.first_bucket))
-		return "offset before first bucket";
-
-	if (bucket_remainder(ca, ptr->offset) +
-	    size_ondisk > ca->mi.bucket_size)
-		return "spans multiple buckets";
-
-	return NULL;
+	return k.k->type == KEY_TYPE_reservation
+		? bkey_s_c_to_reservation(k).v->nr_replicas
+		: bch2_bkey_dirty_devs(k).nr;
 }
 
-const char *bch2_bkey_ptrs_invalid(const struct bch_fs *c, struct bkey_s_c k)
+unsigned bch2_bkey_nr_ptrs_fully_allocated(struct bkey_s_c k)
 {
+	unsigned ret = 0;
+
+	if (k.k->type == KEY_TYPE_reservation) {
+		ret = bkey_s_c_to_reservation(k).v->nr_replicas;
+	} else {
 		struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
 		const union bch_extent_entry *entry;
-	struct bch_extent_crc_unpacked crc;
-	unsigned size_ondisk = k.k->size;
-	const char *reason;
-	unsigned nonce = UINT_MAX;
-
-	if (k.k->type == KEY_TYPE_btree_ptr)
-		size_ondisk = c->opts.btree_node_size;
-
-	bkey_extent_entry_for_each(ptrs, entry) {
-		if (__extent_entry_type(entry) >= BCH_EXTENT_ENTRY_MAX)
-			return "invalid extent entry type";
-
-		if (k.k->type == KEY_TYPE_btree_ptr &&
-		    !extent_entry_is_ptr(entry))
-			return "has non ptr field";
-
-		switch (extent_entry_type(entry)) {
-		case BCH_EXTENT_ENTRY_ptr:
-			reason = extent_ptr_invalid(c, k, &entry->ptr,
-						    size_ondisk, false);
-			if (reason)
-				return reason;
-			break;
-		case BCH_EXTENT_ENTRY_crc32:
-		case BCH_EXTENT_ENTRY_crc64:
-		case BCH_EXTENT_ENTRY_crc128:
-			crc = bch2_extent_crc_unpack(k.k, entry_to_crc(entry));
-
-			if (crc.offset + crc.live_size >
-			    crc.uncompressed_size)
-				return "checksum offset + key size > uncompressed size";
-
-			size_ondisk = crc.compressed_size;
-
-			if (!bch2_checksum_type_valid(c, crc.csum_type))
-				return "invalid checksum type";
-
-			if (crc.compression_type >= BCH_COMPRESSION_NR)
-				return "invalid compression type";
+		struct extent_ptr_decoded p;
 
-			if (bch2_csum_type_is_encryption(crc.csum_type)) {
-				if (nonce == UINT_MAX)
-					nonce = crc.offset + crc.nonce;
-				else if (nonce != crc.offset + crc.nonce)
-					return "incorrect nonce";
-			}
-			break;
-		case BCH_EXTENT_ENTRY_stripe_ptr:
-			break;
-		}
+		bkey_for_each_ptr_decode(k.k, ptrs, p, entry)
+			ret += !p.ptr.cached &&
+				p.crc.compression_type == BCH_COMPRESSION_NONE;
 	}
 
-	return NULL;
-}
-
-/* Btree ptrs */
-
-const char *bch2_btree_ptr_invalid(const struct bch_fs *c, struct bkey_s_c k)
-{
-	if (bkey_val_u64s(k.k) > BKEY_BTREE_PTR_VAL_U64s_MAX)
-		return "value too big";
-
-	return bch2_bkey_ptrs_invalid(c, k);
+	return ret;
 }
 
-void bch2_btree_ptr_debugcheck(struct bch_fs *c, struct bkey_s_c k)
+unsigned bch2_bkey_sectors_compressed(struct bkey_s_c k)
 {
 	struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
-	const struct bch_extent_ptr *ptr;
-	const char *err;
-	char buf[160];
-	struct bucket_mark mark;
-	struct bch_dev *ca;
-
-	bch2_fs_bug_on(!test_bit(BCH_FS_REBUILD_REPLICAS, &c->flags) &&
-		       !bch2_bkey_replicas_marked(c, k, false), c,
-		       "btree key bad (replicas not marked in superblock):\n%s",
-		       (bch2_bkey_val_to_text(&PBUF(buf), c, k), buf));
-
-	if (!test_bit(BCH_FS_INITIAL_GC_DONE, &c->flags))
-		return;
-
-	bkey_for_each_ptr(ptrs, ptr) {
-		ca = bch_dev_bkey_exists(c, ptr->dev);
-
-		mark = ptr_bucket_mark(ca, ptr);
-
-		err = "stale";
-		if (gen_after(mark.gen, ptr->gen))
-			goto err;
-
-		err = "inconsistent";
-		if (mark.data_type != BCH_DATA_BTREE ||
-		    mark.dirty_sectors < c->opts.btree_node_size)
-			goto err;
-	}
+	const union bch_extent_entry *entry;
+	struct extent_ptr_decoded p;
+	unsigned ret = 0;
 
-	return;
-err:
-	bch2_bkey_val_to_text(&PBUF(buf), c, k);
-	bch2_fs_bug(c, "%s btree pointer %s: bucket %zi gen %i mark %08x",
-		    err, buf, PTR_BUCKET_NR(ca, ptr),
-		    mark.gen, (unsigned) mark.v.counter);
-}
+	bkey_for_each_ptr_decode(k.k, ptrs, p, entry)
+		if (!p.ptr.cached &&
+		    p.crc.compression_type != BCH_COMPRESSION_NONE)
+			ret += p.crc.compressed_size;
 
-void bch2_btree_ptr_to_text(struct printbuf *out, struct bch_fs *c,
-			    struct bkey_s_c k)
-{
-	bch2_bkey_ptrs_to_text(out, c, k);
+	return ret;
 }
 
-/* Extents */
-
-int bch2_cut_front_s(struct bpos where, struct bkey_s k)
+bool bch2_check_range_allocated(struct bch_fs *c, struct bpos pos, u64 size,
+				unsigned nr_replicas)
 {
-	unsigned new_val_u64s = bkey_val_u64s(k.k);
-	int val_u64s_delta;
-	u64 sub;
-
-	if (bkey_cmp(where, bkey_start_pos(k.k)) <= 0)
-		return 0;
-
-	EBUG_ON(bkey_cmp(where, k.k->p) > 0);
-
-	sub = where.offset - bkey_start_offset(k.k);
-
-	k.k->size -= sub;
-
-	if (!k.k->size) {
-		k.k->type = KEY_TYPE_deleted;
-		new_val_u64s = 0;
-	}
-
-	switch (k.k->type) {
-	case KEY_TYPE_extent:
-	case KEY_TYPE_reflink_v: {
-		struct bkey_ptrs ptrs = bch2_bkey_ptrs(k);
-		union bch_extent_entry *entry;
-		bool seen_crc = false;
-
-		bkey_extent_entry_for_each(ptrs, entry) {
-			switch (extent_entry_type(entry)) {
-			case BCH_EXTENT_ENTRY_ptr:
-				if (!seen_crc)
-					entry->ptr.offset += sub;
-				break;
-			case BCH_EXTENT_ENTRY_crc32:
-				entry->crc32.offset += sub;
-				break;
-			case BCH_EXTENT_ENTRY_crc64:
-				entry->crc64.offset += sub;
-				break;
-			case BCH_EXTENT_ENTRY_crc128:
-				entry->crc128.offset += sub;
-				break;
-			case BCH_EXTENT_ENTRY_stripe_ptr:
-				break;
-			}
+	struct btree_trans trans;
+	struct btree_iter *iter;
+	struct bpos end = pos;
+	struct bkey_s_c k;
+	bool ret = true;
+	int err;
 
-			if (extent_entry_is_crc(entry))
-				seen_crc = true;
-		}
+	end.offset += size;
 
-		break;
-	}
-	case KEY_TYPE_reflink_p: {
-		struct bkey_s_reflink_p p = bkey_s_to_reflink_p(k);
+	bch2_trans_init(&trans, c, 0, 0);
 
-		le64_add_cpu(&p.v->idx, sub);
+	for_each_btree_key(&trans, iter, BTREE_ID_EXTENTS, pos,
+			   BTREE_ITER_SLOTS, k, err) {
+		if (bkey_cmp(bkey_start_pos(k.k), end) >= 0)
 			break;
-	}
-	case KEY_TYPE_inline_data: {
-		struct bkey_s_inline_data d = bkey_s_to_inline_data(k);
 
-		sub = min_t(u64, sub << 9, bkey_val_bytes(d.k));
-
-		memmove(d.v->data,
-			d.v->data + sub,
-			bkey_val_bytes(d.k) - sub);
-
-		new_val_u64s -= sub >> 3;
+		if (nr_replicas > bch2_bkey_nr_ptrs_fully_allocated(k)) {
+			ret = false;
 			break;
 		}
 	}
+	bch2_trans_exit(&trans);
 
-	val_u64s_delta = bkey_val_u64s(k.k) - new_val_u64s;
-	BUG_ON(val_u64s_delta < 0);
-
-	set_bkey_val_u64s(k.k, new_val_u64s);
-	memset(bkey_val_end(k), 0, val_u64s_delta * sizeof(u64));
-	return -val_u64s_delta;
+	return ret;
 }
 
-int bch2_cut_back_s(struct bpos where, struct bkey_s k)
+static unsigned bch2_extent_ptr_durability(struct bch_fs *c,
+					   struct extent_ptr_decoded p)
 {
-	unsigned new_val_u64s = bkey_val_u64s(k.k);
-	int val_u64s_delta;
-	u64 len = 0;
+	unsigned durability = 0;
+	struct bch_dev *ca;
 
-	if (bkey_cmp(where, k.k->p) >= 0)
+	if (p.ptr.cached)
 		return 0;
 
-	EBUG_ON(bkey_cmp(where, bkey_start_pos(k.k)) < 0);
+	ca = bch_dev_bkey_exists(c, p.ptr.dev);
 
-	len = where.offset - bkey_start_offset(k.k);
+	if (ca->mi.state != BCH_MEMBER_STATE_FAILED)
+		durability = max_t(unsigned, durability, ca->mi.durability);
 
-	k.k->p = where;
-	k.k->size = len;
+	if (p.has_ec) {
+		struct stripe *s =
+			genradix_ptr(&c->stripes[0], p.ec.idx);
 
-	if (!len) {
-		k.k->type = KEY_TYPE_deleted;
-		new_val_u64s = 0;
-	}
+		if (WARN_ON(!s))
+			goto out;
 
-	switch (k.k->type) {
-	case KEY_TYPE_inline_data:
-		new_val_u64s = min(new_val_u64s, k.k->size << 6);
-		break;
+		durability = max_t(unsigned, durability, s->nr_redundant);
 	}
-
-	val_u64s_delta = bkey_val_u64s(k.k) - new_val_u64s;
-	BUG_ON(val_u64s_delta < 0);
-
-	set_bkey_val_u64s(k.k, new_val_u64s);
-	memset(bkey_val_end(k), 0, val_u64s_delta * sizeof(u64));
-	return -val_u64s_delta;
-}
-
-const char *bch2_extent_invalid(const struct bch_fs *c, struct bkey_s_c k)
-{
-	return bch2_bkey_ptrs_invalid(c, k);
+out:
+	return durability;
 }
 
-void bch2_extent_debugcheck(struct bch_fs *c, struct bkey_s_c k)
+unsigned bch2_bkey_durability(struct bch_fs *c, struct bkey_s_c k)
 {
-	struct bkey_s_c_extent e = bkey_s_c_to_extent(k);
+	struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
 	const union bch_extent_entry *entry;
 	struct extent_ptr_decoded p;
-	char buf[160];
-
-	/*
-	 * XXX: we should be doing most/all of these checks at startup time,
-	 * where we check bch2_bkey_invalid() in btree_node_read_done()
-	 *
-	 * But note that we can't check for stale pointers or incorrect gc marks
-	 * until after journal replay is done (it might be an extent that's
-	 * going to get overwritten during replay)
-	 */
-
-	if (percpu_down_read_trylock(&c->mark_lock)) {
-		bch2_fs_bug_on(!test_bit(BCH_FS_REBUILD_REPLICAS, &c->flags) &&
-			       !bch2_bkey_replicas_marked_locked(c, e.s_c, false), c,
-			       "extent key bad (replicas not marked in superblock):\n%s",
-			       (bch2_bkey_val_to_text(&PBUF(buf), c, e.s_c), buf));
-		percpu_up_read(&c->mark_lock);
-	}
-	/*
-	 * If journal replay hasn't finished, we might be seeing keys
-	 * that will be overwritten by the time journal replay is done:
-	 */
-	if (!test_bit(JOURNAL_REPLAY_DONE, &c->journal.flags))
-		return;
-
-	extent_for_each_ptr_decode(e, p, entry) {
-		struct bch_dev *ca	= bch_dev_bkey_exists(c, p.ptr.dev);
-		struct bucket_mark mark = ptr_bucket_mark(ca, &p.ptr);
-		unsigned stale		= gen_after(mark.gen, p.ptr.gen);
-		unsigned disk_sectors	= ptr_disk_sectors(p);
-		unsigned mark_sectors	= p.ptr.cached
-			? mark.cached_sectors
-			: mark.dirty_sectors;
-
-		bch2_fs_bug_on(stale && !p.ptr.cached, c,
-			       "stale dirty pointer (ptr gen %u bucket %u",
-			       p.ptr.gen, mark.gen);
+	unsigned durability = 0;
 
-		bch2_fs_bug_on(stale > 96, c, "key too stale: %i", stale);
+	bkey_for_each_ptr_decode(k.k, ptrs, p, entry)
+		durability += bch2_extent_ptr_durability(c, p);
 
-		bch2_fs_bug_on(!stale &&
-			       (mark.data_type != BCH_DATA_USER ||
-				mark_sectors < disk_sectors), c,
-			       "extent pointer not marked: %s:\n"
-			       "type %u sectors %u < %u",
-			       (bch2_bkey_val_to_text(&PBUF(buf), c, e.s_c), buf),
-			       mark.data_type,
-			       mark_sectors, disk_sectors);
-	}
+	return durability;
 }
 
-void bch2_extent_to_text(struct printbuf *out, struct bch_fs *c,
-			 struct bkey_s_c k)
+void bch2_bkey_mark_replicas_cached(struct bch_fs *c, struct bkey_s k,
+				    unsigned target,
+				    unsigned nr_desired_replicas)
 {
-	bch2_bkey_ptrs_to_text(out, c, k);
-}
+	struct bkey_ptrs ptrs = bch2_bkey_ptrs(k);
+	union bch_extent_entry *entry;
+	struct extent_ptr_decoded p;
+	int extra = bch2_bkey_durability(c, k.s_c) - nr_desired_replicas;
 
-static unsigned bch2_crc_field_size_max[] = {
-	[BCH_EXTENT_ENTRY_crc32] = CRC32_SIZE_MAX,
-	[BCH_EXTENT_ENTRY_crc64] = CRC64_SIZE_MAX,
-	[BCH_EXTENT_ENTRY_crc128] = CRC128_SIZE_MAX,
-};
+	if (target && extra > 0)
+		bkey_for_each_ptr_decode(k.k, ptrs, p, entry) {
+			int n = bch2_extent_ptr_durability(c, p);
 
-static void bch2_extent_crc_pack(union bch_extent_crc *dst,
-				 struct bch_extent_crc_unpacked src,
-				 enum bch_extent_entry_type type)
-{
-#define set_common_fields(_dst, _src)					\
-		_dst.type		= 1 << type;			\
-		_dst.csum_type		= _src.csum_type,		\
-		_dst.compression_type	= _src.compression_type,	\
-		_dst._compressed_size	= _src.compressed_size - 1,	\
-		_dst._uncompressed_size	= _src.uncompressed_size - 1,	\
-		_dst.offset		= _src.offset
+			if (n && n <= extra &&
+			    !bch2_dev_in_target(c, p.ptr.dev, target)) {
+				entry->ptr.cached = true;
+				extra -= n;
+			}
+		}
 
-	switch (type) {
-	case BCH_EXTENT_ENTRY_crc32:
-		set_common_fields(dst->crc32, src);
-		dst->crc32.csum	 = *((__le32 *) &src.csum.lo);
-		break;
-	case BCH_EXTENT_ENTRY_crc64:
-		set_common_fields(dst->crc64, src);
-		dst->crc64.nonce	= src.nonce;
-		dst->crc64.csum_lo	= src.csum.lo;
-		dst->crc64.csum_hi	= *((__le16 *) &src.csum.hi);
-		break;
-	case BCH_EXTENT_ENTRY_crc128:
-		set_common_fields(dst->crc128, src);
-		dst->crc128.nonce	= src.nonce;
-		dst->crc128.csum	= src.csum;
-		break;
-	default:
-		BUG();
+	if (extra > 0)
+		bkey_for_each_ptr_decode(k.k, ptrs, p, entry) {
+			int n = bch2_extent_ptr_durability(c, p);
+
+			if (n && n <= extra) {
+				entry->ptr.cached = true;
+				extra -= n;
+			}
 		}
-#undef set_common_fields
 }
 
-void bch2_extent_crc_append(struct bkey_i *k,
-			    struct bch_extent_crc_unpacked new)
-{
-	struct bkey_ptrs ptrs = bch2_bkey_ptrs(bkey_i_to_s(k));
-	union bch_extent_crc *crc = (void *) ptrs.end;
-	enum bch_extent_entry_type type;
-
-	if (bch_crc_bytes[new.csum_type]	<= 4 &&
-	    new.uncompressed_size - 1		<= CRC32_SIZE_MAX &&
-	    new.nonce				<= CRC32_NONCE_MAX)
-		type = BCH_EXTENT_ENTRY_crc32;
-	else if (bch_crc_bytes[new.csum_type]	<= 10 &&
-		   new.uncompressed_size - 1	<= CRC64_SIZE_MAX &&
-		   new.nonce			<= CRC64_NONCE_MAX)
-		type = BCH_EXTENT_ENTRY_crc64;
-	else if (bch_crc_bytes[new.csum_type]	<= 16 &&
-		   new.uncompressed_size - 1	<= CRC128_SIZE_MAX &&
-		   new.nonce			<= CRC128_NONCE_MAX)
-		type = BCH_EXTENT_ENTRY_crc128;
-	else
-		BUG();
+void bch2_bkey_append_ptr(struct bkey_i *k,
+			  struct bch_extent_ptr ptr)
+{
+	EBUG_ON(bch2_bkey_has_device(bkey_i_to_s_c(k), ptr.dev));
 
-	bch2_extent_crc_pack(crc, new, type);
+	switch (k->k.type) {
+	case KEY_TYPE_btree_ptr:
+	case KEY_TYPE_extent:
+		EBUG_ON(bkey_val_u64s(&k->k) >= BKEY_EXTENT_VAL_U64s_MAX);
 
-	k->k.u64s += extent_entry_u64s(ptrs.end);
+		ptr.type = 1 << BCH_EXTENT_ENTRY_ptr;
 
-	EBUG_ON(bkey_val_u64s(&k->k) > BKEY_EXTENT_VAL_U64s_MAX);
+		memcpy((void *) &k->v + bkey_val_bytes(&k->k),
+		       &ptr,
+		       sizeof(ptr));
+		k->u64s++;
+		break;
+	default:
+		BUG();
+	}
 }
 
 static inline void __extent_entry_insert(struct bkey_i *k,
@@ -1010,6 +799,107 @@ void bch2_extent_ptr_decoded_append(struct bkey_i *k,
 	}
 }
 
+static union bch_extent_entry *extent_entry_prev(struct bkey_ptrs ptrs,
+					  union bch_extent_entry *entry)
+{
+	union bch_extent_entry *i = ptrs.start;
+
+	if (i == entry)
+		return NULL;
+
+	while (extent_entry_next(i) != entry)
+		i = extent_entry_next(i);
+	return i;
+}
+
+union bch_extent_entry *bch2_bkey_drop_ptr(struct bkey_s k,
+					   struct bch_extent_ptr *ptr)
+{
+	struct bkey_ptrs ptrs = bch2_bkey_ptrs(k);
+	union bch_extent_entry *dst, *src, *prev;
+	bool drop_crc = true;
+
+	EBUG_ON(ptr < &ptrs.start->ptr ||
+		ptr >= &ptrs.end->ptr);
+	EBUG_ON(ptr->type != 1 << BCH_EXTENT_ENTRY_ptr);
+
+	src = extent_entry_next(to_entry(ptr));
+	if (src != ptrs.end &&
+	    !extent_entry_is_crc(src))
+		drop_crc = false;
+
+	dst = to_entry(ptr);
+	while ((prev = extent_entry_prev(ptrs, dst))) {
+		if (extent_entry_is_ptr(prev))
+			break;
+
+		if (extent_entry_is_crc(prev)) {
+			if (drop_crc)
+				dst = prev;
+			break;
+		}
+
+		dst = prev;
+	}
+
+	memmove_u64s_down(dst, src,
+			  (u64 *) ptrs.end - (u64 *) src);
+	k.k->u64s -= (u64 *) src - (u64 *) dst;
+
+	return dst;
+}
+
+void bch2_bkey_drop_device(struct bkey_s k, unsigned dev)
+{
+	struct bch_extent_ptr *ptr;
+
+	bch2_bkey_drop_ptrs(k, ptr, ptr->dev == dev);
+}
+
+const struct bch_extent_ptr *
+bch2_bkey_has_device(struct bkey_s_c k, unsigned dev)
+{
+	struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
+	const struct bch_extent_ptr *ptr;
+
+	bkey_for_each_ptr(ptrs, ptr)
+		if (ptr->dev == dev)
+			return ptr;
+
+	return NULL;
+}
+
+bool bch2_bkey_has_target(struct bch_fs *c, struct bkey_s_c k, unsigned target)
+{
+	struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
+	const struct bch_extent_ptr *ptr;
+
+	bkey_for_each_ptr(ptrs, ptr)
+		if (bch2_dev_in_target(c, ptr->dev, target) &&
+		    (!ptr->cached ||
+		     !ptr_stale(bch_dev_bkey_exists(c, ptr->dev), ptr)))
+			return true;
+
+	return false;
+}
+
+bool bch2_bkey_matches_ptr(struct bch_fs *c, struct bkey_s_c k,
+			   struct bch_extent_ptr m, u64 offset)
+{
+	struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
+	const union bch_extent_entry *entry;
+	struct extent_ptr_decoded p;
+
+	bkey_for_each_ptr_decode(k.k, ptrs, p, entry)
+		if (p.ptr.dev	== m.dev &&
+		    p.ptr.gen	== m.gen &&
+		    (s64) p.ptr.offset + p.crc.offset - bkey_start_offset(k.k) ==
+		    (s64) m.offset  - offset)
+			return true;
+
+	return false;
+}
+
 /*
  * bch_extent_normalize - clean up an extent, dropping stale pointers etc.
  *
@@ -1027,245 +917,307 @@ bool bch2_extent_normalize(struct bch_fs *c, struct bkey_s k)
 		ptr_stale(bch_dev_bkey_exists(c, ptr->dev), ptr));
 
 	/* will only happen if all pointers were cached: */
-	if (!bkey_val_u64s(k.k))
+	if (!bch2_bkey_nr_ptrs(k.s_c))
 		k.k->type = KEY_TYPE_discard;
 
 	return bkey_whiteout(k.k);
 }
 
-void bch2_bkey_mark_replicas_cached(struct bch_fs *c, struct bkey_s k,
-				    unsigned target,
-				    unsigned nr_desired_replicas)
+void bch2_bkey_ptrs_to_text(struct printbuf *out, struct bch_fs *c,
+			    struct bkey_s_c k)
 {
-	struct bkey_ptrs ptrs = bch2_bkey_ptrs(k);
-	union bch_extent_entry *entry;
-	struct extent_ptr_decoded p;
-	int extra = bch2_bkey_durability(c, k.s_c) - nr_desired_replicas;
+	struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
+	const union bch_extent_entry *entry;
+	struct bch_extent_crc_unpacked crc;
+	const struct bch_extent_ptr *ptr;
+	const struct bch_extent_stripe_ptr *ec;
+	struct bch_dev *ca;
+	bool first = true;
 
-	if (target && extra > 0)
-		bkey_for_each_ptr_decode(k.k, ptrs, p, entry) {
-			int n = bch2_extent_ptr_durability(c, p);
+	bkey_extent_entry_for_each(ptrs, entry) {
+		if (!first)
+			pr_buf(out, " ");
 
-			if (n && n <= extra &&
-			    !bch2_dev_in_target(c, p.ptr.dev, target)) {
-				entry->ptr.cached = true;
-				extra -= n;
-			}
-		}
+		switch (__extent_entry_type(entry)) {
+		case BCH_EXTENT_ENTRY_ptr:
+			ptr = entry_to_ptr(entry);
+			ca = ptr->dev < c->sb.nr_devices && c->devs[ptr->dev]
+				? bch_dev_bkey_exists(c, ptr->dev)
+				: NULL;
 
-	if (extra > 0)
-		bkey_for_each_ptr_decode(k.k, ptrs, p, entry) {
-			int n = bch2_extent_ptr_durability(c, p);
+			pr_buf(out, "ptr: %u:%llu gen %u%s%s", ptr->dev,
+			       (u64) ptr->offset, ptr->gen,
+			       ptr->cached ? " cached" : "",
+			       ca && ptr_stale(ca, ptr)
+			       ? " stale" : "");
+			break;
+		case BCH_EXTENT_ENTRY_crc32:
+		case BCH_EXTENT_ENTRY_crc64:
+		case BCH_EXTENT_ENTRY_crc128:
+			crc = bch2_extent_crc_unpack(k.k, entry_to_crc(entry));
 
-			if (n && n <= extra) {
-				entry->ptr.cached = true;
-				extra -= n;
+			pr_buf(out, "crc: c_size %u size %u offset %u nonce %u csum %u compress %u",
+			       crc.compressed_size,
+			       crc.uncompressed_size,
+			       crc.offset, crc.nonce,
+			       crc.csum_type,
+			       crc.compression_type);
+			break;
+		case BCH_EXTENT_ENTRY_stripe_ptr:
+			ec = &entry->stripe_ptr;
+
+			pr_buf(out, "ec: idx %llu block %u",
+			       (u64) ec->idx, ec->block);
+			break;
+		default:
+			pr_buf(out, "(invalid extent entry %.16llx)", *((u64 *) entry));
+			return;
 		}
+
+		first = false;
 	}
 }
 
-enum merge_result bch2_extent_merge(struct bch_fs *c,
-				    struct bkey_s _l, struct bkey_s _r)
+static const char *extent_ptr_invalid(const struct bch_fs *c,
+				      struct bkey_s_c k,
+				      const struct bch_extent_ptr *ptr,
+				      unsigned size_ondisk,
+				      bool metadata)
 {
-	struct bkey_s_extent l = bkey_s_to_extent(_l);
-	struct bkey_s_extent r = bkey_s_to_extent(_r);
-	union bch_extent_entry *en_l = l.v->start;
-	union bch_extent_entry *en_r = r.v->start;
-	struct bch_extent_crc_unpacked crc_l, crc_r;
+	struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
+	const struct bch_extent_ptr *ptr2;
+	struct bch_dev *ca;
 
-	if (bkey_val_u64s(l.k) != bkey_val_u64s(r.k))
-		return BCH_MERGE_NOMERGE;
+	if (!bch2_dev_exists2(c, ptr->dev))
+		return "pointer to invalid device";
 
-	crc_l = bch2_extent_crc_unpack(l.k, NULL);
+	ca = bch_dev_bkey_exists(c, ptr->dev);
+	if (!ca)
+		return "pointer to invalid device";
 
-	extent_for_each_entry(l, en_l) {
-		en_r = vstruct_idx(r.v, (u64 *) en_l - l.v->_data);
+	bkey_for_each_ptr(ptrs, ptr2)
+		if (ptr != ptr2 && ptr->dev == ptr2->dev)
+			return "multiple pointers to same device";
 
-		if (extent_entry_type(en_l) != extent_entry_type(en_r))
-			return BCH_MERGE_NOMERGE;
+	if (ptr->offset + size_ondisk > bucket_to_sector(ca, ca->mi.nbuckets))
+		return "offset past end of device";
 
-		switch (extent_entry_type(en_l)) {
-		case BCH_EXTENT_ENTRY_ptr: {
-			const struct bch_extent_ptr *lp = &en_l->ptr;
-			const struct bch_extent_ptr *rp = &en_r->ptr;
-			struct bch_dev *ca;
+	if (ptr->offset < bucket_to_sector(ca, ca->mi.first_bucket))
+		return "offset before first bucket";
 
-			if (lp->offset + crc_l.compressed_size != rp->offset ||
-			    lp->dev			!= rp->dev ||
-			    lp->gen			!= rp->gen)
-				return BCH_MERGE_NOMERGE;
+	if (bucket_remainder(ca, ptr->offset) +
+	    size_ondisk > ca->mi.bucket_size)
+		return "spans multiple buckets";
 
-			/* We don't allow extents to straddle buckets: */
-			ca = bch_dev_bkey_exists(c, lp->dev);
+	return NULL;
+}
 
-			if (PTR_BUCKET_NR(ca, lp) != PTR_BUCKET_NR(ca, rp))
-				return BCH_MERGE_NOMERGE;
+const char *bch2_bkey_ptrs_invalid(const struct bch_fs *c, struct bkey_s_c k)
+{
+	struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
+	const union bch_extent_entry *entry;
+	struct bch_extent_crc_unpacked crc;
+	unsigned size_ondisk = k.k->size;
+	const char *reason;
+	unsigned nonce = UINT_MAX;
+
+	if (k.k->type == KEY_TYPE_btree_ptr)
+		size_ondisk = c->opts.btree_node_size;
+
+	bkey_extent_entry_for_each(ptrs, entry) {
+		if (__extent_entry_type(entry) >= BCH_EXTENT_ENTRY_MAX)
+			return "invalid extent entry type";
 
-			break;
-		}
-		case BCH_EXTENT_ENTRY_stripe_ptr:
-			if (en_l->stripe_ptr.block	!= en_r->stripe_ptr.block ||
-			    en_l->stripe_ptr.idx	!= en_r->stripe_ptr.idx)
-				return BCH_MERGE_NOMERGE;
+		if (k.k->type == KEY_TYPE_btree_ptr &&
+		    !extent_entry_is_ptr(entry))
+			return "has non ptr field";
+
+		switch (extent_entry_type(entry)) {
+		case BCH_EXTENT_ENTRY_ptr:
+			reason = extent_ptr_invalid(c, k, &entry->ptr,
+						    size_ondisk, false);
+			if (reason)
+				return reason;
 			break;
 		case BCH_EXTENT_ENTRY_crc32:
 		case BCH_EXTENT_ENTRY_crc64:
 		case BCH_EXTENT_ENTRY_crc128:
-			crc_l = bch2_extent_crc_unpack(l.k, entry_to_crc(en_l));
-			crc_r = bch2_extent_crc_unpack(r.k, entry_to_crc(en_r));
-
-			if (crc_l.csum_type		!= crc_r.csum_type ||
-			    crc_l.compression_type	!= crc_r.compression_type ||
-			    crc_l.nonce			!= crc_r.nonce)
-				return BCH_MERGE_NOMERGE;
-
-			if (crc_l.offset + crc_l.live_size != crc_l.compressed_size ||
-			    crc_r.offset)
-				return BCH_MERGE_NOMERGE;
+			crc = bch2_extent_crc_unpack(k.k, entry_to_crc(entry));
 
-			if (!bch2_checksum_mergeable(crc_l.csum_type))
-				return BCH_MERGE_NOMERGE;
+			if (crc.offset + crc.live_size >
+			    crc.uncompressed_size)
+				return "checksum offset + key size > uncompressed size";
 
-			if (crc_l.compression_type)
-				return BCH_MERGE_NOMERGE;
+			size_ondisk = crc.compressed_size;
 
-			if (crc_l.csum_type &&
-			    crc_l.uncompressed_size +
-			    crc_r.uncompressed_size > c->sb.encoded_extent_max)
-				return BCH_MERGE_NOMERGE;
+			if (!bch2_checksum_type_valid(c, crc.csum_type))
+				return "invalid checksum type";
 
-			if (crc_l.uncompressed_size + crc_r.uncompressed_size - 1 >
-			    bch2_crc_field_size_max[extent_entry_type(en_l)])
-				return BCH_MERGE_NOMERGE;
+			if (crc.compression_type >= BCH_COMPRESSION_NR)
+				return "invalid compression type";
 
+			if (bch2_csum_type_is_encryption(crc.csum_type)) {
+				if (nonce == UINT_MAX)
+					nonce = crc.offset + crc.nonce;
+				else if (nonce != crc.offset + crc.nonce)
+					return "incorrect nonce";
+			}
+			break;
+		case BCH_EXTENT_ENTRY_stripe_ptr:
 			break;
-		default:
-			return BCH_MERGE_NOMERGE;
 		}
 	}
 
-	extent_for_each_entry(l, en_l) {
-		struct bch_extent_crc_unpacked crc_l, crc_r;
+	return NULL;
+}
 
-		en_r = vstruct_idx(r.v, (u64 *) en_l - l.v->_data);
+void bch2_ptr_swab(const struct bkey_format *f, struct bkey_packed *k)
+{
+	union bch_extent_entry *entry;
+	u64 *d = (u64 *) bkeyp_val(f, k);
+	unsigned i;
 
-		if (!extent_entry_is_crc(en_l))
-			continue;
+	for (i = 0; i < bkeyp_val_u64s(f, k); i++)
+		d[i] = swab64(d[i]);
 
-		crc_l = bch2_extent_crc_unpack(l.k, entry_to_crc(en_l));
-		crc_r = bch2_extent_crc_unpack(r.k, entry_to_crc(en_r));
+	for (entry = (union bch_extent_entry *) d;
+	     entry < (union bch_extent_entry *) (d + bkeyp_val_u64s(f, k));
+	     entry = extent_entry_next(entry)) {
+		switch (extent_entry_type(entry)) {
+		case BCH_EXTENT_ENTRY_ptr:
+			break;
+		case BCH_EXTENT_ENTRY_crc32:
+			entry->crc32.csum = swab32(entry->crc32.csum);
+			break;
+		case BCH_EXTENT_ENTRY_crc64:
+			entry->crc64.csum_hi = swab16(entry->crc64.csum_hi);
+			entry->crc64.csum_lo = swab64(entry->crc64.csum_lo);
+			break;
+		case BCH_EXTENT_ENTRY_crc128:
+			entry->crc128.csum.hi = (__force __le64)
+				swab64((__force u64) entry->crc128.csum.hi);
+			entry->crc128.csum.lo = (__force __le64)
+				swab64((__force u64) entry->crc128.csum.lo);
+			break;
+		case BCH_EXTENT_ENTRY_stripe_ptr:
+			break;
+		}
+	}
+}
 
-		crc_l.csum = bch2_checksum_merge(crc_l.csum_type,
-						 crc_l.csum,
-						 crc_r.csum,
-						 crc_r.uncompressed_size << 9);
+/* Generic extent code: */
 
-		crc_l.uncompressed_size	+= crc_r.uncompressed_size;
-		crc_l.compressed_size	+= crc_r.compressed_size;
+int bch2_cut_front_s(struct bpos where, struct bkey_s k)
+{
+	unsigned new_val_u64s = bkey_val_u64s(k.k);
+	int val_u64s_delta;
+	u64 sub;
 
-		bch2_extent_crc_pack(entry_to_crc(en_l), crc_l,
-				     extent_entry_type(en_l));
-	}
+	if (bkey_cmp(where, bkey_start_pos(k.k)) <= 0)
+		return 0;
 
-	bch2_key_resize(l.k, l.k->size + r.k->size);
+	EBUG_ON(bkey_cmp(where, k.k->p) > 0);
 
-	return BCH_MERGE_MERGE;
-}
+	sub = where.offset - bkey_start_offset(k.k);
 
-bool bch2_check_range_allocated(struct bch_fs *c, struct bpos pos, u64 size,
-			       unsigned nr_replicas)
-{
-	struct btree_trans trans;
-	struct btree_iter *iter;
-	struct bpos end = pos;
-	struct bkey_s_c k;
-	bool ret = true;
-	int err;
+	k.k->size -= sub;
 
-	end.offset += size;
+	if (!k.k->size) {
+		k.k->type = KEY_TYPE_deleted;
+		new_val_u64s = 0;
+	}
 
-	bch2_trans_init(&trans, c, 0, 0);
+	switch (k.k->type) {
+	case KEY_TYPE_extent:
+	case KEY_TYPE_reflink_v: {
+		struct bkey_ptrs ptrs = bch2_bkey_ptrs(k);
+		union bch_extent_entry *entry;
+		bool seen_crc = false;
 
-	for_each_btree_key(&trans, iter, BTREE_ID_EXTENTS, pos,
-			   BTREE_ITER_SLOTS, k, err) {
-		if (bkey_cmp(bkey_start_pos(k.k), end) >= 0)
+		bkey_extent_entry_for_each(ptrs, entry) {
+			switch (extent_entry_type(entry)) {
+			case BCH_EXTENT_ENTRY_ptr:
+				if (!seen_crc)
+					entry->ptr.offset += sub;
 				break;
-
-		if (nr_replicas > bch2_bkey_nr_ptrs_allocated(k)) {
-			ret = false;
+			case BCH_EXTENT_ENTRY_crc32:
+				entry->crc32.offset += sub;
+				break;
+			case BCH_EXTENT_ENTRY_crc64:
+				entry->crc64.offset += sub;
+				break;
+			case BCH_EXTENT_ENTRY_crc128:
+				entry->crc128.offset += sub;
+				break;
+			case BCH_EXTENT_ENTRY_stripe_ptr:
 				break;
 			}
-	}
-	bch2_trans_exit(&trans);
-
-	return ret;
-}
-
-unsigned bch2_bkey_nr_ptrs_allocated(struct bkey_s_c k)
-{
-	unsigned ret = 0;
 
-	switch (k.k->type) {
-	case KEY_TYPE_extent: {
-		struct bkey_s_c_extent e = bkey_s_c_to_extent(k);
-		const union bch_extent_entry *entry;
-		struct extent_ptr_decoded p;
+			if (extent_entry_is_crc(entry))
+				seen_crc = true;
+		}
 
-		extent_for_each_ptr_decode(e, p, entry)
-			ret += !p.ptr.cached &&
-				p.crc.compression_type == BCH_COMPRESSION_NONE;
 		break;
 	}
-	case KEY_TYPE_reservation:
-		ret = bkey_s_c_to_reservation(k).v->nr_replicas;
+	case KEY_TYPE_reflink_p: {
+		struct bkey_s_reflink_p p = bkey_s_to_reflink_p(k);
+
+		le64_add_cpu(&p.v->idx, sub);
 		break;
 	}
+	case KEY_TYPE_inline_data: {
+		struct bkey_s_inline_data d = bkey_s_to_inline_data(k);
 
-	return ret;
-}
-
-/* KEY_TYPE_reservation: */
+		sub = min_t(u64, sub << 9, bkey_val_bytes(d.k));
 
-const char *bch2_reservation_invalid(const struct bch_fs *c, struct bkey_s_c k)
-{
-	struct bkey_s_c_reservation r = bkey_s_c_to_reservation(k);
+		memmove(d.v->data,
+			d.v->data + sub,
+			bkey_val_bytes(d.k) - sub);
 
-	if (bkey_val_bytes(k.k) != sizeof(struct bch_reservation))
-		return "incorrect value size";
+		new_val_u64s -= sub >> 3;
+		break;
+	}
+	}
 
-	if (!r.v->nr_replicas || r.v->nr_replicas > BCH_REPLICAS_MAX)
-		return "invalid nr_replicas";
+	val_u64s_delta = bkey_val_u64s(k.k) - new_val_u64s;
+	BUG_ON(val_u64s_delta < 0);
 
-	return NULL;
+	set_bkey_val_u64s(k.k, new_val_u64s);
+	memset(bkey_val_end(k), 0, val_u64s_delta * sizeof(u64));
+	return -val_u64s_delta;
 }
 
-void bch2_reservation_to_text(struct printbuf *out, struct bch_fs *c,
-			      struct bkey_s_c k)
+int bch2_cut_back_s(struct bpos where, struct bkey_s k)
 {
-	struct bkey_s_c_reservation r = bkey_s_c_to_reservation(k);
+	unsigned new_val_u64s = bkey_val_u64s(k.k);
+	int val_u64s_delta;
+	u64 len = 0;
 
-	pr_buf(out, "generation %u replicas %u",
-	       le32_to_cpu(r.v->generation),
-	       r.v->nr_replicas);
-}
+	if (bkey_cmp(where, k.k->p) >= 0)
+		return 0;
 
-enum merge_result bch2_reservation_merge(struct bch_fs *c,
-					 struct bkey_s _l, struct bkey_s _r)
-{
-	struct bkey_s_reservation l = bkey_s_to_reservation(_l);
-	struct bkey_s_reservation r = bkey_s_to_reservation(_r);
+	EBUG_ON(bkey_cmp(where, bkey_start_pos(k.k)) < 0);
 
-	if (l.v->generation != r.v->generation ||
-	    l.v->nr_replicas != r.v->nr_replicas)
-		return BCH_MERGE_NOMERGE;
+	len = where.offset - bkey_start_offset(k.k);
 
-	if ((u64) l.k->size + r.k->size > KEY_SIZE_MAX) {
-		bch2_key_resize(l.k, KEY_SIZE_MAX);
-		bch2_cut_front_s(l.k->p, r.s);
-		return BCH_MERGE_PARTIAL;
+	k.k->p = where;
+	k.k->size = len;
+
+	if (!len) {
+		k.k->type = KEY_TYPE_deleted;
+		new_val_u64s = 0;
 	}
 
-	bch2_key_resize(l.k, l.k->size + r.k->size);
+	switch (k.k->type) {
+	case KEY_TYPE_inline_data:
+		new_val_u64s = min(new_val_u64s, k.k->size << 6);
+		break;
+	}
 
-	return BCH_MERGE_MERGE;
+	val_u64s_delta = bkey_val_u64s(k.k) - new_val_u64s;
+	BUG_ON(val_u64s_delta < 0);
+
+	set_bkey_val_u64s(k.k, new_val_u64s);
+	memset(bkey_val_end(k), 0, val_u64s_delta * sizeof(u64));
+	return -val_u64s_delta;
 }

fs/bcachefs/extents.h

@@ -40,6 +40,9 @@ struct btree_insert_entry;
 		(union bch_extent_entry *) (_entry));			\
 })
 
+#define extent_entry_next(_entry)					\
+	((typeof(_entry)) ((void *) (_entry) + extent_entry_bytes(_entry)))
+
 static inline unsigned
 __extent_entry_type(const union bch_extent_entry *e)
 {
@@ -185,10 +188,52 @@ struct bkey_ptrs {
 	union bch_extent_entry	*end;
 };
 
-/* iterate over bkey ptrs */
+static inline struct bkey_ptrs_c bch2_bkey_ptrs_c(struct bkey_s_c k)
+{
+	switch (k.k->type) {
+	case KEY_TYPE_btree_ptr: {
+		struct bkey_s_c_btree_ptr e = bkey_s_c_to_btree_ptr(k);
+		return (struct bkey_ptrs_c) {
+			to_entry(&e.v->start[0]),
+			to_entry(extent_entry_last(e))
+		};
+	}
+	case KEY_TYPE_extent: {
+		struct bkey_s_c_extent e = bkey_s_c_to_extent(k);
+		return (struct bkey_ptrs_c) {
+			e.v->start,
+			extent_entry_last(e)
+		};
+	}
+	case KEY_TYPE_stripe: {
+		struct bkey_s_c_stripe s = bkey_s_c_to_stripe(k);
+		return (struct bkey_ptrs_c) {
+			to_entry(&s.v->ptrs[0]),
+			to_entry(&s.v->ptrs[s.v->nr_blocks]),
+		};
+	}
+	case KEY_TYPE_reflink_v: {
+		struct bkey_s_c_reflink_v r = bkey_s_c_to_reflink_v(k);
 
-#define extent_entry_next(_entry)					\
-	((typeof(_entry)) ((void *) (_entry) + extent_entry_bytes(_entry)))
+		return (struct bkey_ptrs_c) {
+			r.v->start,
+			bkey_val_end(r),
+		};
+	}
+	default:
+		return (struct bkey_ptrs_c) { NULL, NULL };
+	}
+}
+
+static inline struct bkey_ptrs bch2_bkey_ptrs(struct bkey_s k)
+{
+	struct bkey_ptrs_c p = bch2_bkey_ptrs_c(k.s_c);
+
+	return (struct bkey_ptrs) {
+		(void *) p.start,
+		(void *) p.end
+	};
+}
 
 #define __bkey_extent_entry_for_each_from(_start, _end, _entry)		\
 	for ((_entry) = (_start);					\
@@ -281,96 +326,26 @@ out:									\
 #define bkey_for_each_crc(_k, _p, _crc, _iter)				\
 	__bkey_for_each_crc(_k, (_p).start, (_p).end, _crc, _iter)
 
-/* utility code common to all keys with pointers: */
-
-static inline struct bkey_ptrs_c bch2_bkey_ptrs_c(struct bkey_s_c k)
-{
-	switch (k.k->type) {
-	case KEY_TYPE_btree_ptr: {
-		struct bkey_s_c_btree_ptr e = bkey_s_c_to_btree_ptr(k);
-		return (struct bkey_ptrs_c) {
-			to_entry(&e.v->start[0]),
-			to_entry(extent_entry_last(e))
-		};
-	}
-	case KEY_TYPE_extent: {
-		struct bkey_s_c_extent e = bkey_s_c_to_extent(k);
-		return (struct bkey_ptrs_c) {
-			e.v->start,
-			extent_entry_last(e)
-		};
-	}
-	case KEY_TYPE_stripe: {
-		struct bkey_s_c_stripe s = bkey_s_c_to_stripe(k);
-		return (struct bkey_ptrs_c) {
-			to_entry(&s.v->ptrs[0]),
-			to_entry(&s.v->ptrs[s.v->nr_blocks]),
-		};
-	}
-	case KEY_TYPE_reflink_v: {
-		struct bkey_s_c_reflink_v r = bkey_s_c_to_reflink_v(k);
-
-		return (struct bkey_ptrs_c) {
-			r.v->start,
-			bkey_val_end(r),
-		};
-	}
-	default:
-		return (struct bkey_ptrs_c) { NULL, NULL };
-	}
-}
-
-static inline struct bkey_ptrs bch2_bkey_ptrs(struct bkey_s k)
-{
-	struct bkey_ptrs_c p = bch2_bkey_ptrs_c(k.s_c);
+/* Iterate over pointers in KEY_TYPE_extent: */
 
-	return (struct bkey_ptrs) {
-		(void *) p.start,
-		(void *) p.end
-	};
-}
-
-static inline struct bch_devs_list bch2_bkey_devs(struct bkey_s_c k)
-{
-	struct bch_devs_list ret = (struct bch_devs_list) { 0 };
-	struct bkey_ptrs_c p = bch2_bkey_ptrs_c(k);
-	const struct bch_extent_ptr *ptr;
-
-	bkey_for_each_ptr(p, ptr)
-		ret.devs[ret.nr++] = ptr->dev;
-
-	return ret;
-}
-
-static inline struct bch_devs_list bch2_bkey_dirty_devs(struct bkey_s_c k)
-{
-	struct bch_devs_list ret = (struct bch_devs_list) { 0 };
-	struct bkey_ptrs_c p = bch2_bkey_ptrs_c(k);
-	const struct bch_extent_ptr *ptr;
-
-	bkey_for_each_ptr(p, ptr)
-		if (!ptr->cached)
-			ret.devs[ret.nr++] = ptr->dev;
+#define extent_for_each_entry_from(_e, _entry, _start)			\
+	__bkey_extent_entry_for_each_from(_start,			\
+				extent_entry_last(_e),_entry)
 
-	return ret;
-}
+#define extent_for_each_entry(_e, _entry)				\
+	extent_for_each_entry_from(_e, _entry, (_e).v->start)
 
-static inline struct bch_devs_list bch2_bkey_cached_devs(struct bkey_s_c k)
-{
-	struct bch_devs_list ret = (struct bch_devs_list) { 0 };
-	struct bkey_ptrs_c p = bch2_bkey_ptrs_c(k);
-	const struct bch_extent_ptr *ptr;
+#define extent_ptr_next(_e, _ptr)					\
+	__bkey_ptr_next(_ptr, extent_entry_last(_e))
 
-	bkey_for_each_ptr(p, ptr)
-		if (ptr->cached)
-			ret.devs[ret.nr++] = ptr->dev;
+#define extent_for_each_ptr(_e, _ptr)					\
+	__bkey_for_each_ptr(&(_e).v->start->ptr, extent_entry_last(_e), _ptr)
 
-	return ret;
-}
+#define extent_for_each_ptr_decode(_e, _ptr, _entry)			\
+	__bkey_for_each_ptr_decode((_e).k, (_e).v->start,		\
+				   extent_entry_last(_e), _ptr, _entry)
 
-unsigned bch2_bkey_nr_ptrs(struct bkey_s_c);
-unsigned bch2_bkey_nr_dirty_ptrs(struct bkey_s_c);
-unsigned bch2_bkey_durability(struct bch_fs *, struct bkey_s_c);
+/* utility code common to all keys with pointers: */
 
 void bch2_mark_io_failure(struct bch_io_failures *,
 			  struct extent_ptr_decoded *);
@@ -378,22 +353,12 @@ int bch2_bkey_pick_read_device(struct bch_fs *, struct bkey_s_c,
 			       struct bch_io_failures *,
 			       struct extent_ptr_decoded *);
 
-void bch2_bkey_append_ptr(struct bkey_i *, struct bch_extent_ptr);
-void bch2_bkey_drop_device(struct bkey_s, unsigned);
-const struct bch_extent_ptr *bch2_bkey_has_device(struct bkey_s_c, unsigned);
-bool bch2_bkey_has_target(struct bch_fs *, struct bkey_s_c, unsigned);
-
-void bch2_bkey_ptrs_to_text(struct printbuf *, struct bch_fs *,
-			    struct bkey_s_c);
-const char *bch2_bkey_ptrs_invalid(const struct bch_fs *, struct bkey_s_c);
-
-/* bch_btree_ptr: */
+/* KEY_TYPE_btree_ptr: */
 
 const char *bch2_btree_ptr_invalid(const struct bch_fs *, struct bkey_s_c);
 void bch2_btree_ptr_debugcheck(struct bch_fs *, struct bkey_s_c);
 void bch2_btree_ptr_to_text(struct printbuf *, struct bch_fs *,
 			    struct bkey_s_c);
-void bch2_ptr_swab(const struct bkey_format *, struct bkey_packed *);
 
 #define bch2_bkey_ops_btree_ptr (struct bkey_ops) {		\
 	.key_invalid	= bch2_btree_ptr_invalid,		\
@@ -402,12 +367,11 @@ void bch2_ptr_swab(const struct bkey_format *, struct bkey_packed *);
 	.swab		= bch2_ptr_swab,			\
 }
 
-/* bch_extent: */
+/* KEY_TYPE_extent: */
 
 const char *bch2_extent_invalid(const struct bch_fs *, struct bkey_s_c);
 void bch2_extent_debugcheck(struct bch_fs *, struct bkey_s_c);
 void bch2_extent_to_text(struct printbuf *, struct bch_fs *, struct bkey_s_c);
-bool bch2_extent_normalize(struct bch_fs *, struct bkey_s);
 enum merge_result bch2_extent_merge(struct bch_fs *,
 				    struct bkey_s, struct bkey_s);
 
@@ -420,7 +384,7 @@ enum merge_result bch2_extent_merge(struct bch_fs *,
 	.key_merge	= bch2_extent_merge,			\
 }
 
-/* bch_reservation: */
+/* KEY_TYPE_reservation: */
 
 const char *bch2_reservation_invalid(const struct bch_fs *, struct bkey_s_c);
 void bch2_reservation_to_text(struct printbuf *, struct bch_fs *, struct bkey_s_c);
@@ -433,13 +397,15 @@ enum merge_result bch2_reservation_merge(struct bch_fs *,
 	.key_merge	= bch2_reservation_merge,		\
 }
 
-void bch2_bkey_mark_replicas_cached(struct bch_fs *, struct bkey_s,
-				    unsigned, unsigned);
+/* Extent checksum entries: */
 
-unsigned bch2_extent_is_compressed(struct bkey_s_c);
+bool bch2_can_narrow_extent_crcs(struct bkey_s_c,
+				 struct bch_extent_crc_unpacked);
+bool bch2_bkey_narrow_crcs(struct bkey_i *, struct bch_extent_crc_unpacked);
+void bch2_extent_crc_append(struct bkey_i *,
+			    struct bch_extent_crc_unpacked);
 
-bool bch2_bkey_matches_ptr(struct bch_fs *, struct bkey_s_c,
-			   struct bch_extent_ptr, u64);
+/* Generic code for keys with pointers: */
 
 static inline bool bkey_extent_is_direct_data(const struct bkey *k)
 {
@@ -477,34 +443,57 @@ static inline bool bkey_extent_is_allocation(const struct bkey *k)
 	}
 }
 
-/* Extent entry iteration: */
+static inline struct bch_devs_list bch2_bkey_devs(struct bkey_s_c k)
+{
+	struct bch_devs_list ret = (struct bch_devs_list) { 0 };
+	struct bkey_ptrs_c p = bch2_bkey_ptrs_c(k);
+	const struct bch_extent_ptr *ptr;
 
-#define extent_for_each_entry_from(_e, _entry, _start)			\
-	__bkey_extent_entry_for_each_from(_start,			\
-				extent_entry_last(_e),_entry)
+	bkey_for_each_ptr(p, ptr)
+		ret.devs[ret.nr++] = ptr->dev;
 
-#define extent_for_each_entry(_e, _entry)				\
-	extent_for_each_entry_from(_e, _entry, (_e).v->start)
+	return ret;
+}
 
-#define extent_ptr_next(_e, _ptr)					\
-	__bkey_ptr_next(_ptr, extent_entry_last(_e))
+static inline struct bch_devs_list bch2_bkey_dirty_devs(struct bkey_s_c k)
+{
+	struct bch_devs_list ret = (struct bch_devs_list) { 0 };
+	struct bkey_ptrs_c p = bch2_bkey_ptrs_c(k);
+	const struct bch_extent_ptr *ptr;
 
-#define extent_for_each_ptr(_e, _ptr)					\
-	__bkey_for_each_ptr(&(_e).v->start->ptr, extent_entry_last(_e), _ptr)
+	bkey_for_each_ptr(p, ptr)
+		if (!ptr->cached)
+			ret.devs[ret.nr++] = ptr->dev;
 
-#define extent_for_each_ptr_decode(_e, _ptr, _entry)			\
-	__bkey_for_each_ptr_decode((_e).k, (_e).v->start,		\
-				   extent_entry_last(_e), _ptr, _entry)
+	return ret;
+}
 
-void bch2_extent_crc_append(struct bkey_i *,
-			    struct bch_extent_crc_unpacked);
-void bch2_extent_ptr_decoded_append(struct bkey_i *,
-				    struct extent_ptr_decoded *);
+static inline struct bch_devs_list bch2_bkey_cached_devs(struct bkey_s_c k)
+{
+	struct bch_devs_list ret = (struct bch_devs_list) { 0 };
+	struct bkey_ptrs_c p = bch2_bkey_ptrs_c(k);
+	const struct bch_extent_ptr *ptr;
 
-bool bch2_can_narrow_extent_crcs(struct bkey_s_c,
-				 struct bch_extent_crc_unpacked);
-bool bch2_bkey_narrow_crcs(struct bkey_i *, struct bch_extent_crc_unpacked);
+	bkey_for_each_ptr(p, ptr)
+		if (ptr->cached)
+			ret.devs[ret.nr++] = ptr->dev;
 
+	return ret;
+}
+
+unsigned bch2_bkey_nr_ptrs(struct bkey_s_c);
+unsigned bch2_bkey_nr_ptrs_allocated(struct bkey_s_c);
+unsigned bch2_bkey_nr_ptrs_fully_allocated(struct bkey_s_c);
+unsigned bch2_bkey_sectors_compressed(struct bkey_s_c);
+bool bch2_check_range_allocated(struct bch_fs *, struct bpos, u64, unsigned);
+unsigned bch2_bkey_durability(struct bch_fs *, struct bkey_s_c);
+
+void bch2_bkey_mark_replicas_cached(struct bch_fs *, struct bkey_s,
+				    unsigned, unsigned);
+
+void bch2_bkey_append_ptr(struct bkey_i *, struct bch_extent_ptr);
+void bch2_extent_ptr_decoded_append(struct bkey_i *,
+				    struct extent_ptr_decoded *);
 union bch_extent_entry *bch2_bkey_drop_ptr(struct bkey_s,
 					   struct bch_extent_ptr *);
 
@@ -525,6 +514,22 @@ do {									\
 	}								\
 } while (0)
 
+void bch2_bkey_drop_device(struct bkey_s, unsigned);
+const struct bch_extent_ptr *bch2_bkey_has_device(struct bkey_s_c, unsigned);
+bool bch2_bkey_has_target(struct bch_fs *, struct bkey_s_c, unsigned);
+
+bool bch2_bkey_matches_ptr(struct bch_fs *, struct bkey_s_c,
+			   struct bch_extent_ptr, u64);
+
+bool bch2_extent_normalize(struct bch_fs *, struct bkey_s);
+void bch2_bkey_ptrs_to_text(struct printbuf *, struct bch_fs *,
+			    struct bkey_s_c);
+const char *bch2_bkey_ptrs_invalid(const struct bch_fs *, struct bkey_s_c);
+
+void bch2_ptr_swab(const struct bkey_format *, struct bkey_packed *);
+
+/* Generic extent code: */
+
 int bch2_cut_front_s(struct bpos, struct bkey_s);
 int bch2_cut_back_s(struct bpos, struct bkey_s);
 
@@ -568,7 +573,4 @@ static inline void extent_save(struct btree *b, struct bkey_packed *dst,
 		BUG_ON(!bch2_bkey_pack_key(dst, src, f));
 }
 
-bool bch2_check_range_allocated(struct bch_fs *, struct bpos, u64, unsigned);
-unsigned bch2_bkey_nr_ptrs_allocated(struct bkey_s_c);
-
 #endif /* _BCACHEFS_EXTENTS_H */

fs/bcachefs/fs-io.c

@@ -675,7 +675,7 @@ static void bch2_add_page_sectors(struct bio *bio, struct bkey_s_c k)
 	struct bvec_iter iter;
 	struct bio_vec bv;
 	unsigned nr_ptrs = k.k->type == KEY_TYPE_reflink_v
-		? 0 : bch2_bkey_nr_ptrs_allocated(k);
+		? 0 : bch2_bkey_nr_ptrs_fully_allocated(k);
 	unsigned state = k.k->type == KEY_TYPE_reservation
 		? SECTOR_RESERVED
 		: SECTOR_ALLOCATED;
@@ -2543,7 +2543,7 @@ static long bchfs_fcollapse_finsert(struct bch_inode_info *inode,
 		} else {
 			/* We might end up splitting compressed extents: */
 			unsigned nr_ptrs =
-				bch2_bkey_nr_dirty_ptrs(bkey_i_to_s_c(copy.k));
+				bch2_bkey_nr_ptrs_allocated(bkey_i_to_s_c(copy.k));
 
 			ret = bch2_disk_reservation_get(c, &disk_res,
 					copy.k->k.size, nr_ptrs,
@@ -2669,7 +2669,7 @@ static long bchfs_fallocate(struct bch_inode_info *inode, int mode,
 		bch2_cut_back(end_pos,		&reservation.k_i);
 
 		sectors = reservation.k.size;
-		reservation.v.nr_replicas = bch2_bkey_nr_dirty_ptrs(k);
+		reservation.v.nr_replicas = bch2_bkey_nr_ptrs_allocated(k);
 
 		if (!bkey_extent_is_allocation(k.k)) {
 			ret = bch2_quota_reservation_add(c, inode,
@@ -2680,7 +2680,7 @@ static long bchfs_fallocate(struct bch_inode_info *inode, int mode,
 		}
 
 		if (reservation.v.nr_replicas < replicas ||
-		    bch2_extent_is_compressed(k)) {
+		    bch2_bkey_sectors_compressed(k)) {
 			ret = bch2_disk_reservation_get(c, &disk_res, sectors,
 							replicas, 0);
 			if (unlikely(ret))

fs/bcachefs/io.c

@@ -202,8 +202,8 @@ static int sum_sector_overwrites(struct btree_trans *trans,
 
 	for_each_btree_key_continue(iter, BTREE_ITER_SLOTS, old, ret) {
 		if (!may_allocate &&
-		    bch2_bkey_nr_ptrs_allocated(old) <
-		    bch2_bkey_nr_dirty_ptrs(bkey_i_to_s_c(new))) {
+		    bch2_bkey_nr_ptrs_fully_allocated(old) <
+		    bch2_bkey_nr_ptrs_allocated(bkey_i_to_s_c(new))) {
 			ret = -ENOSPC;
 			break;
 		}

fs/bcachefs/move.c

@@ -134,11 +134,11 @@ static int bch2_migrate_index_update(struct bch_write_op *op)
 		 * If we're not fully overwriting @k, and it's compressed, we
 		 * need a reservation for all the pointers in @insert
 		 */
-		nr = bch2_bkey_nr_dirty_ptrs(bkey_i_to_s_c(insert)) -
+		nr = bch2_bkey_nr_ptrs_allocated(bkey_i_to_s_c(insert)) -
 			 m->nr_ptrs_reserved;
 
 		if (insert->k.size < k.k->size &&
-		    bch2_extent_is_compressed(k) &&
+		    bch2_bkey_sectors_compressed(k) &&
 		    nr > 0) {
 			ret = bch2_disk_reservation_add(c, &op->res,
 					keylist_sectors(keys) * nr, 0);
@@ -250,7 +250,7 @@ int bch2_migrate_write_init(struct bch_fs *c, struct migrate_write *m,
 		 */
 #if 0
 		int nr = (int) io_opts.data_replicas -
-			bch2_bkey_nr_dirty_ptrs(k);
+			bch2_bkey_nr_ptrs_allocated(k);
 #endif
 		int nr = (int) io_opts.data_replicas;
 
@@ -599,7 +599,7 @@ static int __bch2_move_data(struct bch_fs *c,
 		if (rate)
 			bch2_ratelimit_increment(rate, k.k->size);
 next:
-		atomic64_add(k.k->size * bch2_bkey_nr_dirty_ptrs(k),
+		atomic64_add(k.k->size * bch2_bkey_nr_ptrs_allocated(k),
 			     &stats->sectors_seen);
 next_nondata:
 		bch2_btree_iter_next(iter);

fs/bcachefs/recovery.c

@@ -254,7 +254,7 @@ static int bch2_extent_replay_key(struct bch_fs *c, enum btree_id btree_id,
 	 * Some extents aren't equivalent - w.r.t. what the triggers do
 	 * - if they're split:
 	 */
-	bool remark_if_split = bch2_extent_is_compressed(bkey_i_to_s_c(k)) ||
+	bool remark_if_split = bch2_bkey_sectors_compressed(bkey_i_to_s_c(k)) ||
 		k->k.type == KEY_TYPE_reflink_p;
 	bool remark = false;
 	int ret;
@@ -289,7 +289,7 @@ static int bch2_extent_replay_key(struct bch_fs *c, enum btree_id btree_id,
 		    bkey_cmp(atomic_end, k->k.p) < 0) {
 			ret = bch2_disk_reservation_add(c, &disk_res,
 					k->k.size *
-					bch2_bkey_nr_dirty_ptrs(bkey_i_to_s_c(k)),
+					bch2_bkey_nr_ptrs_allocated(bkey_i_to_s_c(k)),
 					BCH_DISK_RESERVATION_NOFAIL);
 			BUG_ON(ret);